Adversarial spheres

Pytorch implementation Adversarial Spheres.

Requiresments

python 3.6.0
pytorch 1.4.0
numpy 1.15.0
tensorflow 2.1.0
tqdm

How to run

• regular training:

$ python main.py --method clean

• train with the exact maximizer of the inner-max optimization (truemax)

$ python main.py --method truemax

• adversarial training with PGD attacks

$ python main.py --method adv --pgd_alpha 0.01 --pgd_itr 100

Standard training

Training with trueax

Adversarial training with PGD examples (eps = 0.01, itr = 100)

Some minor typos in paper

Here I notice a few typos in the original paper. They are very minor and will not affect the overal understanding of the paper, but they do matter for reproducibility.

• In Equation 3, Analytic error rate on the inner sphere should be:
  

• Right above Equation 3, the variance is missing a square sign:
  

• The Gaussian distribution in the caption of Figure F5 should be:
  

Discussions

Number of iterations required for a perfect classifier

Others

Please cite the following paper for Adversarial spheres:

@inproceedings{46623,
title	= {Adversarial Spheres},
author	= {Justin Gilmer and Luke Metz and Fartash Faghri and Sam Schoenholz and Maithra Raghu and Martin Wattenberg and Ian Goodfellow},
year	= {2018},
URL	= {https://arxiv.org/pdf/1801.02774.pdf}
}

Contact

If you have any questions or suggestions, please feel free to contact me via ama at cs dot toronto dot edu!